Bit Office gives AI automation a visible, controllable workspace. Different AI models collaborate as one team under a Team Leader — planning, coding, reviewing, and delivering in a single flow, all rendered in a live pixel-art office you can watch, control, and share.
What makes it different: agents get better over time. Rate each project on creativity, visual quality, interaction, completeness, and engagement. Your ratings and review patterns are stored as persistent memory — the next time the team plans a project, they know what scored low and actively improve on it.
Auto-detect installed AI CLIs (Claude, Codex, Gemini, Aider, OpenCode)
Generate a pair code for mobile access
Features
Multi-Agent Teams
A Team Leader coordinates specialists like Developer and Code
Reviewer to plan, implement, and validate tasks automatically.
Multi-Model Workflows
Run Claude, Codex, Gemini, Aider, and OpenCode together in one
pipeline, letting each model focus on what it does best.
Pixel Office Workspace
Watch agents work in real time inside a PixiJS pixel-art office,
with live status, logs, and progress visualization.
Instant Preview & Rating
Every completed task generates an auto preview. Rate the result
across five dimensions — your feedback becomes persistent memory
that shapes how agents approach the next project.
Self-Improving Agents
Review patterns, tech preferences, and project ratings are stored
across sessions. Agents learn what you value and adapt — low
visual scores lead to richer designs, recurring review failures get
avoided automatically.
Token Cost Visibility
Track token usage per agent and per team in real time so you always
know the cost of each run.
Live Sharing & Mobile Control
Invite others to watch progress, leave feedback, or manage sessions
directly from your phone.
Cross-Device Sync
Real-time collaboration powered by WebSocket, Ably, and Telegram
channels.
Project History
Every run is saved with a replayable preview, letting you revisit
results and build on previous work.
Team Workflow
| Phase | What Happens | Your Action |
|---|---|---|
| Create | Team Lead gathers intent and scope | Describe what to build |
| Design | Team Lead proposes implementation plan | Approve or request changes |
| Execute | Developer implements, Reviewer validates | Monitor or cancel |
| Complete | Preview and summary delivered | Rate, give feedback, or end project |
Ratings persist as agent memory. The next project starts with lessons from the last one.
AI-native prototyping — go from idea to working preview in one session
Feature spikes — rapid implementation with continuous preview feedback
Multi-model experiments — compare how different AI backends approach the same task
Live demos — show autonomous development workflows to your team or audience
Run from Source
Prerequisites
Node.js 18+
pnpm
At least one AI CLI installed: claude, codex, gemini, aider, or opencode
Setup
git clone https://github.com/longyangxi/bit-office.git
cd bit-office
pnpm install
pnpm dev
Scripts
| Command | Description |
|---|---|
| pnpm dev | Web + gateway in dev mode |
| pnpm dev:web | Web only (Next.js) |
| pnpm dev:gateway | Gateway only |
| pnpm dev:desktop | Tauri desktop app (dev mode) |
| pnpm build | Build all packages |
| pnpm build:desktop | Build Tauri .app + .dmg |
| pnpm start | Build web and start gateway |
Environment Variables
| Variable | Required | Description |
|---|---|---|
| WORKSPACE | No | Agent working directory |
| ABLY_API_KEY | No | Remote real-time channel |
| TELEGRAM_BOT_TOKENS | No | One token per bot/agent (comma-separated) |
| WS_PORT | No | Gateway WebSocket port (default: 9090) |
| WEB_DIR | No | Override served web build directory |
Desktop App (Tauri)
Bit Office also ships as a native macOS desktop app powered by Tauri. The app bundles the gateway as a sidecar — no terminal, no browser, just launch and go.
The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
